These smaller types give you a memory reduction only when you store them in unboxed arrays or similar. There, each will take as many bits as indicated by the type suffix.
In general use, they all take exactly as much storage as an Int or Word, the main difference is that the values are automatically narrowed to the appropriate bit size when using fixed-width types, and there are (still) more optimisations (in the form of rewrite rules mainly) for Int and Word than for Int8 etc., so some operations will be slower using those.
Concerning the question whether to use Word64 throughout or to use smaller types, that depends. On a 64-bit system, when compiling with optimisations, the performance of Word and Word64 should mostly be the same since where it matters both should be unpacked and the work is done on the raw machine Word#. But there probably still are a few rules for Word that have no Word64 counterpart yet, so perhaps there is a difference after all. On a 32-bit system, most operations on Word64 are implemented via C calls, so there operations on Word64 are much slower than operations on Word.
So depending on what is more important, simplicity of code or performance on different systems, either
- use
Word64 throughout: simple code, good performance on 64-bit systems
- use
Word as long as your values are guaranteed to fit into 32 bits and transform to Word64 at the latest safe moment: more complicated code, but better performance on 32-bit systems.